Dr. Long Cheng · Senior Architect at NVIDIA
I build TensorRT-LLM infrastructure for NVIDIA's latest GPUs, spanning sparse attention, lossless Top-K, and LLM for system and algorithm design and optimization, from GPU kernels to compiler workflows such as LLM4Compiler.
My work sits at the boundary between algorithms, systems, and hardware. I focus on building efficient AI infrastructure that turns research ideas into production-quality implementations with measurable impact.
Before NVIDIA, I worked at Huawei Beijing Research Center on low-precision LLM operators, heterogeneous acceleration, and LLM-native algorithm and compiler workflows. My background also includes high-performance scientific computing, numerical methods, and mixed-precision algorithms from research at Beihang University and the University of Cambridge.
Industry roles and research engineering experience.
Senior Architect, Compute DL Architecture
Building TensorRT-LLM inference infrastructure for the latest NVIDIA GPUs, with focus on sparse attention, LLM-native algorithm design, GPU kernel optimization, long-context decoding, and system-algorithm co-design.
Researcher to Principal Research Engineer, Compilers and Programming Languages Lab
Led low-precision LLM operators, heterogeneous acceleration, and LLM-native compiler workflows, including PASA, LLM4Compiler, VecTrans, and mixed-precision solver optimization across Ascend and Kunpeng.
PhD Intern Researcher
Researched scientific computing and mixed-precision methods for PDE workloads in collaboration with UK-based teams.
Academic training in engineering, mathematics, and scientific computing.
PhD, Aerospace Engineering · Supervisor: Prof. Xiaofeng Sun
PhD research in computational aeroacoustics, immersed boundary methods, mixed precision, and massively parallel scientific computing.
CSC-Sponsored Full-time Visiting PhD Student, Computational and Applied Mathematics · Host Supervisor: Prof. Paul Tucker
Worked on high-performance scientific computing under a full-time visiting PhD arrangement during the doctoral program.
BEng, Aerospace Engineering
Representative projects and technical artifacts.
A four-stage, lossless Top-K algorithm and GPU kernel for DSA sparse attention: search-based, sort-free, and almost atomic-free, designed for long-context inference and integrated into TensorRT-LLM.
Developed through AI-assisted, human-in-the-loop kernel co-design, with AI accelerating algorithm exploration and optimization while human judgment drove problem framing, correctness, and final performance decisions.
LLM-native compiler and code optimization initiative including VecTrans and G2CTrans for transformation and performance-oriented workflows.
Robust low-precision attention algorithm for large-model inference on Ascend NPUs, combining algorithmic, numerical, and systems considerations.
Massively parallel CFD/CAA solver for flow-sound interaction in turbomachinery, forming the foundation of later work in numerical methods and performance engineering.
Selected external talks and organizational recognition.
Contact
The best way to reach me is by email. You can also find my code, publications, and professional profile through the links below.